Professional Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Keywords: This paper discusses a hypothetical case study in which SPRAY and CALPUFF dispersion models are applied to
Atmospheric dispersion modelling the simulation of an incidental fire. For this type of accident, source features are typically not directly
Sensitivity analysis measurable, thus making their definition critical. The choice of some model-specific parameters is another
Environmental impact
critical issue, since clear indications are rarely available in guidelines. The aim of this work is to compare how
Models comparison
Source characterization
pollutant concentrations simulated with the two models are affected by changing these two sets of data (i.e.
Fire simulation parameters related to the emission source and model specific parameters), thus performing a sensitivity study to
identify the most influential variables. The most relevant outcome is that sensitivities of the two models are
generally comparable, except for the source diameter: if the SPRAY model is applied with the specific fire source
option, then the concentrations result almost independent from this parameter. Conversely, when considering
other source-types, the concentrations vary up to ± 60% within the selected uncertainty range.
Peer review under responsibility of Turkish National Committee for Air Pollution Research and Control.
* Corresponding author.
E-mail address: marzio.invernizzi@polimi.it (M. Invernizzi).
https://doi.org/10.1016/j.apr.2021.101249
Received 6 June 2021; Received in revised form 26 October 2021; Accepted 26 October 2021
Available online 3 November 2021
1309-1042/© 2021 Turkish National Committee for Air Pollution Research and Control. Production and hosting by Elsevier B.V. All rights reserved.
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
(Holnicki and Nahorski, 2015). In view of this, a sensitivity analysis is parameters cannot be generalized to any model or to any situation, but it
important to explore and quantify the impact of possible changes in refers specifically to the selected case-study. This work does not aim to
input data on the model outputs. suggest how to perform a sensitivity study applicable to any situation,
The aim of this paper is, based on a hypothetical case-study of a nor addresses model developers interested in developing or upgrading
refinery fire, to discuss the sensitivity of two dispersion models applied software. Actually, it highlights the importance of investigating the
to simulate the dispersion of atmospheric pollutants, with the purpose of possible range of variation of the input data to identify the most influ
investigating the most influential model input data. ential variables, and it may represent a sort of guideline for model users
As hypothetical case-study, we decided to consider a relatively small who have to deal with the implementation of similar case-studies.
fire, involving a portion of gas oil treatment unit. The reason for this Even though, in recent years, the use of dispersion models to simu
choice is that the application of dispersion models is consistent in such late pollutant dispersion into the atmosphere is continuously increasing,
conditions, whereas for large catastrophic explosions the lack of repre to the best of our knowledge, in literature there are few studies relevant
sentation of the explosive phase or plume buoyancy in many models to model comparison in terms of sensitivity (Antonioni et al., 2012;
limits their application. Björnham et al., 2020; Devenish et al., 2012; Gant et al., 2013; Srinivas
This work discusses the influence of possible errors in the source data et al., 2016). Also, the novelty of this work is related to the approach
and in model-specific parameters on the results. The first objective is followed to evaluate the model sensitivity. Indeed, the paper is not
particularly interesting because of the high uncertainty in the estimation limited to investigate the influence of the source term parameters on the
of source term parameters for fires (Daly et al., 2012). Indeed, in the case model outputs, but it also considers model-specific parameters. The
of fires, source geometrical features are hardly directly measurable, but latter are particularly interesting because for many of them, clear in
they have to be estimated using specific correlations. In our case study, dications on their setting are lacking both in the literature and in the
the source term parameters (i.e. height, diameter and temperature) were model user’s guides, thus making it a critical issue for the modelist.
varied within a fairly limited range, since, as previously mentioned, we Moreover, differently from other papers, this work proposes a novel dual
decided to simulate a localized fire involving a single equipment. approach to address the issues that are relevant to the definition of the
Furthermore, the choice of some model-specific parameters is often input data by investigating the model sensitivity on two different levels.
critical, because clear indications are rarely available in literature or in
the specific model user’s guides, and so their definition is generally left 2. Materials and methods
to the professional judgement of the model user.
The work described in this paper uses two different approaches to 2.1. Case – study description
assess the variability in the model results. First, starting from a “base-
case”, defined as the most representative scenario for the case study, The hypothesized case study regards an incidental fire in an oil re
different “plausible” emissive scenarios are investigated. Each of these finery: the event is supposed to involve a portion of the gas oil treatment
scenarios is characterised by a “macroscopic variation” of a single unit.
parameter within the hypothesized uncertainty range. In this study, the In our hypothesis, the fire lasts 3 h. In real cases, the duration is a
term “macroscopic variation” refers to the extremes of the considered fundamental point to define the case study and must be evaluated based
range of variation of each parameter. The results obtained for each on the statements of people who were on the spot.
scenario are compared to the results of the “base-case”. This way, it is To optimize the choice of the geographic simulation domain, the
possible to evaluate the effect caused by a wrong estimation of an input plume direction should be considered. For the selected case study, a
datum, which is crucially important especially in the case of environ rectangular domain of 25 × 25 km has been identified with a mesh grid
mental and health impact assessments. This first approach allows to of 250 m. Then, assuming that the plume evolves in south-western di
identify the most influential parameters on the model outputs. rection, the source has been located at the north-eastern corner of the
The second approach aims to further analyse the effect of these pa domain.
rameters by considering “microscopic” variations thereof, i.e. a small In addition, for a more precise analysis, some discrete receptors
Thus, this second investigation is intended to evaluate effectively the should be positioned in places considered of particular interest to esti
model sensitivity to the different parameters, by imposing the same mate the pollutants concentration resulting from the incidental fire (e.g.,
small percent perturbation (compared to the reference base-case) to all, hospitals, schools, city hall). In the identified scenario some receptors
regardless of the reasonable range of variability. have been located as to be representative of possible places of interest.
However, it should be highlighted that this paper does not aim to
present an exhaustive numerical sensitivity study. The adopted 2.2. Selection of the dispersion model
approach highlights the importance of having a broader view of the
issue relevant to the model sensitivity, thereby carrying out additional A key element for an effective dispersion modelling study is to
investigations focused on those parameters that potentially generate the choose an appropriate modelling tool to match the scale of impact and
highest variations in the results. the complexity of the emission scenario. The choice of the most suitable
It is worth noting that this sensitivity study has not been validated model can be based on the study of the scientific literature and the
with experimental field measurements, because it is a hypothetical analysis of the technical legislation, which are both useful in order to
event. Conversely, in real conditions, a model validation is strongly understand the features of each model.
suggested to evaluate the model accuracy and to select properly the The scientific literature is quite deficient in terms of studies con
reference base-case as the most representative scenario. cerning accidental fires in oil refineries. Also, even though there are
For the purposes of the study, two models have been selected, i.e. the some studies comparing Lagrangian particle models and puff models
Lagrangian particle model SPRAY and the puff model CALPUFF (Elbir, (Invernizzi et al., 2020; Ravina et al., 2020; Souto et al., 2001), to the
2003; Elbir et al., 2010; Holnicki et al., 2016; Rzeszutek, 2019). Indeed, best of our knowledge, papers comparing their sensitivity to input data
Lagrangian particle models and puff models currently represent the most are lacking.
common tools to simulate pollutants dispersion from fires (Adame et al., Nonetheless, there are several reviews (Holmes and Morawska,
2018; Ainslie and Jackson, 2009; Henderson et al., 2008). These models 2006; Leelőssy et al., 2014; Sinha et al., 2004), which illustrate, in a
have been chosen because there is a large variety of studies that prove rather comprehensive way, the different types of models that can be
their validity (Invernizzi et al., 2021), although, as previously used for the simulation of pollutants dispersion. Those reviews typically
mentioned, there are no specific indications on a regulatory level. describe the model’s features and discuss advantages and disadvantages
It should be pointed out that the investigation of the most influential according to the kind of application.
2
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
In particular, according to the literature, Gaussian plume models the simulated area is characterized by a flat land.
(Islam, 1999; Liu et al., 2015), Lagrangian puff models (Jung et al.,
2003), Lagrangian particle models (Cécé et al., 2016; Santiago and 2.4. Definition of base-case
Martín, 2008), Eulerian grid models (Kota et al., 2013; Seland and
Iversen, 1999) and fluid dynamics models (Leelőssy et al., 2014; Mar Before starting a modelling study, it is necessary to quantify some
katos et al., 2009) may be used to simulate pollutants dispersion. A parameters needed as input data.
detailed description of the different classes of dispersion models avail In the case of a “real fire”, the first step is the quantification of the
able is out of the scope of this paper. amount of fuel burnt. This can be estimated by a mass balance around
Here, some considerations have been made in order to choose the the involved equipment.
most suitable model for the selected case study: Then, for an incidental fire, the involved compounds and their
emission factors shall be defined. The scientific and technical literature
• The greatest advantage of Gaussian models is that they have an states that the most common pollutants associated with gas oil com
extremely fast, almost immediate, response time. However, the use bustion are CO2, CO, generic unburnt hydrocarbons (CH), particulate
of very simple models, such as Gaussians, is not advisable in case of matter (PM), SOX and NOX (Booher and Janke, 1997; Lemieux et al.,
large simulation domains (Daly et al., 2012), because they consider 2004).
steady state conditions. Thus, they cannot adequately describe the The SFPE Handbook of Fire Protection Engineering (DiNenno et al.,
dispersive phenomenon, since the meteorological condition of one 2002) has been chosen for the estimation of emission factors, because it
point of the domain is not representative of the wind field variations has been considered to be an authoritative and reliable reference. More
over the entire domain. in detail, Tables 3–4.14 “Yields of Fire Products and Chemical,
• Eulerian and fluid dynamics models are very advanced simulators, Convective, and Radiative Heats of Combustion for Well-Ventilated
but at the same time they are very complex and require a long Fires” has been considered for the definition of the emission factors of
computational time. the CO2, CO, unburnt hydrocarbons (CH) and particulate matter (PM)
• Fluid dynamics models fare suggested, or even necessary, when the emitted during the combustion of different fuels. Among all the listed
dispersion occurs in urban areas, where the influence of buildings on species, the generic “Hydrocarbon” and the kerosene have been
the dispersion is dominant, or where the scale to consider is the so- considered, thus obtaining the following emission factors (ton/ton): 2.7
called meteorological microscale (<1 km) (Bhuiyan and Naser, for CO2, 0.02 for CO, 0.007 for CH, and 0.05 for PM.
2015; Jang et al., 2015; Mishra and Wehrstedt, 2015; Novozhilov, The SFPE Handbook of Fire Protection Engineering does not report
2001). However, the considered case study has a simulation domain emission factors for NOX emissions. Therefore, the emission factor for
larger than 1 km and the simulated site is not located in an urban NOx, which was set equal to 0.01 (ton/ton), was defined using the AP-42
area, making such complex models unneeded. database of the US EPA, 1998.
For SOX we assumed the complete (stoichiometric) conversion to SO2
For these reasons, a Lagrangian puff model (i.e. CALPUFF) and a of the elemental sulphur (S) in the original fuel, which was hypothesized
Lagrangian particle model (i.e. SPRAY) have been chosen: they repre to be 100 ppm, thus giving an emission factor of 0.002 (ton/ton).
sent a compromise between reasonable accuracy and manageable The emission rate for each compound can be calculated as the
computational time. product of its emission factor and the amount of fuel burnt, divided by
Furthermore, the use of these models for the selected case study is the event duration.
compliant with the Italian technical standards on the matter (UNI Furthermore, the model requires the geometrical features of the
10796, 2009; UNI 10964, 2009). These standards define the scenarios simulated source. Therefore, diameter and height have to be evaluated
for the implementation of different models, suggesting the best model from surveys and technical documentation.
for each situation. In addition to the geometrical characterization of the source, some
Finally, the available scientific literature also supports this choice: physical parameters have to be defined. The fire smoke rise velocity is
past studies prove the suitability of these models in similar cases. For derived from the Ingason correlation (Ingason and Li, 2015):
instance, there are some examples of studies carried out using puff
⎛ ⎞13
models, and specifically CALPUFF, for the simulation of pollutant
/
On the other hand, there are fewer articles regarding the application
of SPRAY to the simulation of fires, presumably because it is a more
where Ta is the ambient temperature, ρ0 the air density, cp the specific
recently developed model. However, the bibliographic research proves
heat of air at constant pressure, g is the acceleration of gravity, q the heat
the suitability of the Lagrangian models to simulate fires: the model
release and D the source diameter.
behaviour is usually compared with measured data to test its goodness,
The Ingason correlation requires the estimation of the heat release,
showing in most cases a good accuracy (Adame et al., 2018; Liu et al.,
which may be derived from the Babrauskas correlation (DiNenno et al.,
2018).
2002):
( )
2.3. Meteorological and orographical features of the model q = m′′∞ ⋅ΔhC,eff 1 − exp− (kβD) A (2)
The meteorological data used for the simulations are three- where Δhc,eff is the net heat of combustion, A is the source area, D is the
dimensional prognostic WRF data purchased from Lakes Environ diameter, and kβ and m′′∞ are empirical constants available in literature
mental. Each model processes the WRF data using the model-specific for a number of common fuels. The values for the fuel of interest are
meteorological tools (i.e. SWIFT for SPRAY and CALMET for CAL obtained from DiNenno et al. (2002), Tables 3–1.13 “Pool Burning:
PUFF), which are diagnostic “mass consistent” models. They generate Thermochemical and Empirical Constants for a Number of Common
3D wind fields inside the meteorological domain, which has been set Organic Fuels”, which reports the empirical constants for the most
equal to the computational grid. For the simulations, the meteorological common organic fuels. Since specific data for gas oil are not available,
conditions were chosen to be representative of a neutral winter morning the fuel that has been considered most representative among those listed
with cloud cover. is JP-5 (Δhc,eff = 43 MJ/kg; kβ = 1.6 m− 1; m′′∞ = 0.054 kg/(m2s)).
The model also requires the site orography as input data. In this case, Finally, it is necessary to define the fire temperature. To do this, the
3
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
hydrocarbon fire curve, reported in BS EN 1363-2, 1999 showing the where ε = 0.7 represents the reduction term due to radiation, P is the
trend of the temperature as a function of time, has been used. energy/time, g the acceleration of gravity, cp and ρair the specific heat
The maximum achievable temperature of about 1100 ◦ C, has been
and the density of the ambient air, and Ta is the ambient temperature.
considered for the hypothesized case study since it is rapidly achieved
after a few minutes.
Based on these evaluations, a reference “base-case” was defined by 2.6. Definition of alternative cases for the evaluation of the effects of
setting the input parameters considered as most representative of the source characterization
hypothesized emission scenario.
Its characteristic parameters are shown in Table 1. Starting from the “base-case”, it has been decided to investigate
alternative emission scenarios by changing the most critical geometrical
2.5. Source types and plume rise computation parameters of the source within a reasonable range of variation. Indeed,
the estimation of these variables is characterized by high uncertainty
One of the investigated parameters is the modelled source type. For because of the impossibility of measuring them directly during the
both CALPUFF and SPRAY, the fire is modelled by applying the specific event.
source type suggested for fires (i.e. the buoyant area source for CAL In addition to the source geometrical parameters, other variables
PUFF, and the fire for SPRAY) and then compared with the model results associated with the definition of the emission scenario have been
obtained by assimilating the fire to a point source (i.e. stack). investigated (e.g., temperature and amount of fuel burnt).
According to the CALPUFF User’s Guide (Scire et al., 2000), the Under real conditions of a refinery fire, in order to define the alter
buoyant area source is the most appropriate tool to simulate fires. native scenarios, it is advisable to first validate the base-case, with the
Therefore, for the CALPUFF simulations, the point source has been purpose to verify that this is effectively the most representative and
considered only for the “base-case”, because of its unsuitability to model realistic scenario. However, it shall be considered that this is usually
the selected event. Despite the indications of the User’s Guide, the point very complex in case of accidental fires.
source has been implemented in the CALPUFF “base-case” to highlight The alternative scenarios defined for the study are shown in Table 2,
the differences in the ground concentrations in comparison to the with the numbers in bold representing the variables changed in the
buoyant area source, as will be discussed in Paragraph 3.1, and, alternative scenarios.
consequently, to underline the importance of using the specific fire
options available in the models. 2.7. Evaluation of the effects of model-specific parameters
For each type of source, the dispersion model simulates the plume
rise mechanism according to a different scheme. When considering the 2.7.1. CALPUFF
fire as a point source (without any fire-specific option), both CALPUFF
As far as CALPUFF is concerned, the only model-specific parameter
and SPRAY calculate the buoyancy flux according to the Briggs equation that has been investigated is the initial vertical dispersion coefficient σz0,
(Tinarelli, 2017), which is proportional to the square of the source
which defines the initial dimension of the puff in the vertical direction.
radius: Concerning the base-case scenario, as suggested in the user’s guide for
T − Ta Gaussian plume models (US EPA, 1995), σz0 is evaluated as follows:
Fb = gr2 w0 (3)
Ta
σ z0 = H/2.15 (6)
where g is the acceleration of gravity, r the source radius, w0 the effluent
where H is the source height.
exit velocity, Ta the ambient temperature and T the exit smoke tem
This correlation is recommended for surface-based sources, as it is
perature.
the case in the modelled case-study.
When using the CALPUFF - buoyant area source model, the radiative
However, in literature, other correlations are reported, which can be
heat loss from the plume to the ambient air can be estimated using the
adopted depending on the source elevation or the presence of adjacent
following equation (Scire et al., 2000):
buildings. More in detail, in case of elevated source located on or
/
q 2 ( ) adjacent to a building, it is suggested to divide the building height by a
r = − 2εσ r T 4 − Ta4 cp (4)
cp factor of 2.15. In case of an elevated source not adjacent to a building, it
is recommended to divide the vertical dimension of the source by 4.3
where q is the radiative heat loss, cp the specific heat of the ambient air, r (US EPA, 1995). Accordingly, to run the alternative scenario, σz0 is set
the source radius, σ the Stefan-Boltzmann constant, ε the emissivity, T equal to the source height divided by 4.3.
the plume temperature and Ta the ambient temperature. Here, an in
crease of the radius implies a reduction of the heat losses. Consequently, 2.7.2. SPRAY
the plume rise increases and the pollutant concentrations decrease. SPRAY is a more advanced software: it requires the definition of
On the other hand, if the SPRAY model is used in combination with several parameters, whose estimation is not trivial, due to the absence of
the specific fire option, the equation used for the buoyancy calculation, specific indications.
in which neither velocity nor radius appear, results in a buoyancy flux
not affected by the source diameter (Tinarelli, 2017): 2.7.2.1. Height of the first layer. Differently from CALPUFF, which
gP provides the output concentration in a gridded surface at the ground, the
Fb = ε (5) Lagrangian particle model needs the height of the first layer as input
πcp ρair Ta
datum: it is the height of the first cell above ground used by the model to
Table 1
Source term parameters for the base-case. D = source diameter; T = plume temperature; H = source height; v = exit velocity; q = heat release; Quantity = amount of
fuel burnt.
Scenario D T H v Q Quantity PM CO CO2 NOx SO2 HC
[m] [K] [m] [m/s] [kW] [ton] [g/s] [g/s] [g/s] [g/s] [g/s] [g/s]
BASE 5 1373 15 8.16 46,354 11.2 51.8 20.7 2800 10.4 0.2 7.3
4
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Table 2
Alternative scenarios for the source term parameters.
Scenario D T H v Quantity PM CO CO2 NOx SO2 HC
[m] [K] [m] [m/s] [ton] [g/s] [g/s] [g/s] [g/s] [g/s] [g/s]
It is worth underlining that, when using the SPRAY - fire model, scenarios T1 and T2 have not been considered, since the fire temperature is not an input parameter
required by the software.
compute the concentration in any point P (x; y; z; t) at time t. To do this, area (building downwash) resulting in high concentrations immediately
the model computes the pollutant concentration considering a “sam downwind of the source. Therefore, to reproduce the emission region, a
pling volume” having the grid step dimensions in x and y direction, and conservative value equal to twice the source height (30 m) has been
the third dimension z is the height of the first layer. defined.
The value attributed to this parameter for the base-case is 10 m: on
one hand, in order to evaluate the ground level concentration, the first 2.7.2.3. Number of particles. Another investigated parameter is related
layer height should be sufficiently low and, on the other hand, a too low to the stochastic description of the Lagrangian model. The air is
value is not advisable in order to limit the influence of the mechanical described as a set of parcels that move according to two different
turbulence, the effect of which is more pronounced in the vicinity of the mechanisms: advection and turbulent motion. The irregular and highly
terrain, progressively decreasing moving far from the ground. variable nature of the main parameters describing the motion of a
In addition, the surface roughness shall be considered for the eval molecule in the air makes it not possible to use exact values of them in
uation of this parameter. The roughness length z0 represents the height any practical problem, since a fully deterministic approach is almost
where the wind speed becomes zero (no-slip condition) and it is related impossible.
to the terrain features: depending on the land use type, different values The SPRAY model requires the definition of the number of particles
for z0 are suggested. In particular, in case of urban areas, SPRAY used for the simulations; whereby each particle represents a discrete
(ARIANET, 2011) sets this parameter to 1 m. Hanna and Britter (2002) amount of pollutant. For this variable, a suitable value has been selected
suggested that the ratio between z0 and the obstacle height Hr can be in retrospect, after running some simulations with different numbers of
estimated according to a simple rule of thumb: particles. To identify a reasonable value, a compromise between good
z0 accuracy in the results and manageable computational time has been
= 0.1 (7) considered, leading to a choice of about 3 million of particles emitted in
Hr
the 3 h of simulation. To ensure a detailed description of the particles
It follows that the parameter Hr assumes a value of 10 m. Also, Hanna motion, a sufficiently high number of particles is required. Indeed, the
and Britter (2002) highlighted that for a typical urban or industrial site smaller the sample size is, the more outliers may skew the findings. In
an average building height of 10 m is a reasonable estimation. Since one other words, the particles represent the air parcels. Thus, each particle
of the main purposes of an atmospheric dispersion model is the assess has a random contribution of motion that has to be described consid
ment of the impact on people, it is reasonable to consider the height up ering a stochastic approach. If the particles number is high enough, an
to which the concentration estimation is of interest (i.e. the height of the average behaviour can be identified, minimizing the discrete contribu
first layer) of 10 m. tion of each particle. Conversely, if few particles are considered, there is
the risk that the outliers are heavier. In other words, the number of
2.7.2.2. Δz. Another model-specific parameter that has to be defined is particles should be high enough as to ensure that the results do not show
Δz, i.e. the vertical dimension of the “emission parallelepiped”. SPRAY any statistically meaningful difference when changing this variable.
generates particles uniformly distributed on a “terrain following” Table 3 shows the model-specific parameters for SPRAY and
parallelepiped centred in P (X0, Y0, Z0), which are the coordinates of the
emission region centre of gravity, the vertical dimension of which is Δz
(Tinarelli, 2017). Table 3
SPRAY and CALPUFF model specific parameters for base-case and alternative
In other words, this parallelepiped can be thought of as a box in
scenarios.
which the particles initially appear. Thus, they are released in a vertical
region ranging from Z0– Δz/2 and Z0 + Δz/2, with Z0 coincident with SPRAY CALPUFF
the source height. In the Supplementary Material S1 a sketch of the Scenario Δz H1 Layer
◦ Particle number Scenario σZ0
emission parallelepiped is shown, highlighting its vertical dimension Δz, [m] [m] [m]
and its position with respect to the emission source. BASE 30 10 2′ 989′ 440 BASE 6.98 ( = H/2.15)
This variable describes the initial condition of the emission and it Δz2 15 10 2′ 989′ 440 σZ0,1 3.49 ( = H/4.3)
should be defined as to reproduce the geometrical features of the Δz3 20 10 2′ 989′ 440
25 10 2′ 989′ 440
emission region. The “geometrical features of the emission region” refers
Δz4
H_1L 30 4 2′ 989′ 440
not only to the effective source dimensions, but also to the dynamic PARTICLE1 30 10 151′200
effects affecting the emission. For instance, because of the configuration PARTICLE2 30 10 749′520
of the stack or of the adjacent buildings, the plume may not rise freely in PARTICLE3 30 10 2′492′640
PARTICLE4 30 10 2′989′440
the atmosphere: some aerodynamic effects due to the way the wind
PARTICLE5 30 10 7′471′440
moves around adjacent buildings and the stack can force the plume to PARTICLE6 30 10 10′670′400
wards the ground instead of allowing it to rise. Depending on the stack PARTICLE7 30 10 14′938′560
height, it may be possible for the plume to be pulled down into this wake PARTICLE8 30 10 24′896′160
5
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
CALPUFF selected for the base-case and the alternative scenarios, with both directions with respect to the value of the base-case, while all the
the numbers in bold representing the variables changed in the alterna other parameters were kept unchanged.
tive scenarios. Then, the concentration values resulting from these perturbations
In order to investigate the effect of the particles number, the case have to be computed on a set of discrete receptors. In this case, 40 re
with the highest number of particles (i.e. 74′ 682′ 000 particles emitted in ceptors placed along the plume axis starting from the source and spaced
3 h, from the SPRAY default release option of 34’575 particles in 5 s) has 250 m from each other were considered.
been considered as a reference, whereas the “base-case” with 2′ 989′ 440 This sensitivity study has a different purpose compared to the
particles is treated as an alternative scenario. This is the reason why, in approach based on the evaluation of the alternative scenarios. In this
Table 3, the scenario with 74′ 682′ 000 is not reported, because it is case, the perturbations imposed to the investigated variables are not the
considered as the reference scenario to compute the sensitivity of the extremes of a reasonable range of variability for the selected parameter,
results with respect to the number of particles in the alternative sce but a “microscopic” perturbation is considered. As a result, the sensi
narios (PARTICLE1 - PARTICLE8). tivity analysis is intended to evaluate the numerical model sensitivity to
the investigated variables, regardless of the range of variability thereof.
2.8. Sensitivity analysis In particular, this approach allows to:
Once the most influential parameters have been identified by means • Compare effectively the model sensitivity to different parameters, by
of the simulations relevant to the alternative scenarios, a sensitivity imposing the same perturbations to all of them;
analysis is needed in order to quantify the effects of a “microscopic • Investigate the model behaviour caused by a perturbation of the
variation” of the input variables on the model outputs. selected parameter, identifying, for instance, the relationship exist
ing between the input and the output;
2.8.1. Choice of the approach • Identify the way the model sensitivity to the selected parameter
The choice of the method to be used for the sensitivity analysis was changes if moving far from the source;
based on a deep literature search. As a result, a paper proposing a • Compare SPRAY and CALPUFF sensitivity to the selected parameters,
sensitivity analysis based on the Taylor series approach by Yegnan et al. highlighting which of them is more sensitive to perturbations of the
(2002) has been selected, because it well-suits the hypothesized investigated input data.
case-study.
Indeed, Yegnan et al. (2002) adopted this approach to calculate the One important preliminary consideration concerns the applicability
sensitivity of ground level concentrations resulting from a short-term of this approach to the SPRAY model, whereby the number of particles
simulation (1 h), as in the case ofour hypothesis. Also, after the defini may affect the significance of the calculated sensitivity indexes. Indeed,
tion of a “base-case” for the modelling scenario, each of the seven input to obtain reliable results from this test, since the variation applied to the
parameters (many of which are the same investigated in the alternative parameters is significantly low (i.e. 1%–3%), it is important to reduce as
scenarios reported in Table 2, e.g., stack height, stack diameter, tem much as possible the influence related to the choice of the number of
perature) has been modified to determine the ones that mostly affect the particles.
output concentrations, i.e. the parameters to which the model is most For this reason, it is necessary to adopt a significantly high number of
sensitive. particles to run these simulations. For the selected case-study, Supple
In the paper by Yegnan et al. (2002), the “base-case” identified is the mentary Material S2 shows that simulations with 106′ 688′ 880 particles
most representative scenario the average values are derived from. Then, have negligible variability due to the number of particles. For this
the most sensitive input parameters, i.e. wind speed and temperature, reason, the sensitivity analysis for the SPRAY model is carried out
are perturbed by 1% in both directions with respect to the average considering the emission of 106′ 688′ 880 particles.
values, and the corresponding change in the output is evaluated. The
sensitivity of the output f’(x) is then computed as: 3. Results and critical discussion
′
f (x) =
f (x2 ) − f (x1 )
(8) 3.1. Results of “base-case” for different source types
x2 − x1
The simulation results can be processed with the post processing
where x is the value of the input parameter adopted in the reference
tools of each of the models under investigation (i.e. CALPOST for CAL
base-case, x1 and x2 are the perturbed input values on either side of x ,
PUFF and POSTBIN for SPRAY). This way, ground level concentration
and f(x1 ) and f(x2 ) are the corresponding output values.
maps representing the pollutants dispersion within the simulation
Also, to normalize the results by removing the effects of units, the
domain are obtained. As an example of the simulation results, Fig. 1
dimensionless sensitivity index (Gonsamo, 2011; Rodrigues et al., 2013)
shows the maximum 1-h concentration maps for the PM resulting from
is introduced:
the base-case simulation in function of the different source types
x explained in section 2.5. The simulation of PM dispersion has been run
(9)
′ ′
f (x)N = f (x)⋅
y as for an inert gas (no deposition considered). PM was selected as the
“tracer” to be treated as reference species for further evaluations, since it
where the apex N refers to the normalization of the sensitivity index, and is, among the compounds considered, the one with the highest emission
y is the value of the output resulting from the reference base-case (with x rate after CO2. Therefore, from this point forward, PM will be referred to
as input variable). as “Tracer”. It should be highlighted that the reference compound cho
Rodrigues et al. (2013) defined the normalized sensitivity index of a sen doesn’t affect the forthcoming considerations about model sensi
variable with respect to a parameter as “the ratio of the relative change tivity or the comparison between different source types: the same
in the variable to the relative change in the parameter”. Thus, it is ob considerations could be applied to any other species considered.
tained by multiplying f (x) by the ratio of the parameter value to the Depending on the source type and the dispersion model used, slightly
′
model result for the base-case scenario. different plume directions can be observed. The maps in Fig. 1 differ in
concentration and plume shapes: this can be explained considering the
2.8.2. Application of the method to the case study different wind fields resulting from the application of the different met
To perform the sensitivity study discussed in the previous paragraph, pre-processors. Indeed, despite starting from the same raw meteoro
the investigated parameters have been perturbed by 1%, 2% and 3% in logical input data, the different meteorological pre-processors (i.e.
6
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Fig. 1. Maximum ground level concentration maps of the tracer species resulting from CALPUFF-buoyant area (a.), SPRAY- fire (b.), CALPUFF-point (c.) and SPRAY-
point (d.). Figure (a.) shows the position of the six discrete receptors discussed below.
CALMET for CALPUFF and SWIFT for SPRAY) elaborate the results in a of places of public interest and receptor 6 is the point of maximum
slightly different way. As an example, the picture of the wind field concentration at the plant fence line.
computed by CALMET and SWIFT for the same hour is reported in the Fig. 2 shows the trend of the of maximum 1-h concentration of the
Supplementary Material S3. Tracer as a function of the distance from the source. To the purpose, 40
The maximum Tracer concentrations calculated by the models on a receptors placed along the plume axis starting from the source and
set of selected discrete receptors (indicated in Fig. 1(a.)) are reported in spaced 250 m from each other are considered.
Table 4. More in detail, receptor 1 corresponds to the gridded receptor The different trends shown in Fig. 2 can be explained by considering
where the maximum concentration has been calculated by the two the different plume rise computations for point sources and for fires/
models inside the simulation domain, receptors (2–5) are representative buoyant area sources. According to the SPRAY model for fires, which
considers a non-complete combustion, there is a cold fraction of parti
cles that remains unburnt and immediately falls to the ground, without
Table 4 being dragged into the plume rise. This gives the highest Tracer
Maximum Tracer concentration values at selected receptors calculated by
CALPUFF (left) and SPRAY (right).
ID Description CALPUFF SPRAY
7
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
8
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Fig. 3. % variation of the maximum PM concentration values at the selected receptors resulting from the simulations of the alternative emission scenarios compared
to the reference base-case for the different source types.
9
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Table 5 observed for the change of the vertical dimension of the “emission
% variation of the maximum Tracer concentration values at the selected re parallelepiped”: for the scenarios Δz2, Δz3 and Δz4, the SPRAY - point
ceptors resulting from the A1 and A2 scenarios compared to the reference base- source model always shows a percent variation higher than the fire
case for the different combinations of models and source types considered. model. Thus, this behaviour does not seem to be attributable to the
ID CALPUFF (Buoyant Area) SPRAY (Point) SPRAY (Fire) position of the selected receptors but it is a general feature resulting
A1 A2 A1 A2 A1 A2 from the different way to model the source. Indeed, a change in Δz,
which represents the vertical dimension of a “box” centred in the
1 − 48.10% 42.10% 52.40% 47.60% 10.20% − 0.80%
emission region centre of gravity in which the particles initially appear
−
2 − 61.90% 7.60% − 24.80% 8.10% 2.00% − 2.40%
3 − 40.50% 35.70% − 64.50% 50.90% 2.50% 0.20% (see Section 2.7.2.2), means a change in the dimension of the region
4 32.90% − 53.30% − 39.00% 24.70% 2.10% 1.00% from which the particles start to rise up due to the buoyancy. Thus, the
5 − 67.10% 74.40% − 38.60% 23.20% 1.10% 0.30% plume rise is affected by this variable in the sense that the “idealized”
6 − 45.70% 21.00% 59.90% 49.90% 1.30% 0.30%
−
plume containing the particles has a different initial shape and dimen
sion according to this parameter. On the other hand, the fire model
the case for the variations applied to the source diameter or the amount considers a portion of emitted particles with no buoyancy flux: the rapid
of fuel. downfall of these particles makes this kind of simulated source less
Indeed, both the SPRAY – point source model and the SPRAY - fire dependent on the variable Δz.
model show, in correspondence of almost all the receptors considered Another investigated parameter is the number of particles, which
and for almost all the investigated parameters, a variability lower than determines the variabilities (%) reported in Fig. 4. It is worth recalling
10%. In addition, comparing the two source types, their response to the that, when analysing the output variations due to the particles number,
input variation is very similar. The most different behaviour is the one the variabilities are not referred to the base-case, but to a different
reference simulation involving the highest amount of particles
Fig. 4. % variation of the maximum Tracer concentration values at the selected receptors resulting from the simulations of the alternative cases for the particles
number (for SPRAY – point and SPRAY – fire) compared to the reference base-case.
10
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
considered (i.e. 74′ 682′ 000 particles emitted in 3 h). The higher the simulated by SPRAY than for the CALPUFF model. As discussed in
number of particles, the lower the stochastic variability of the results. Section 3.1, the simulations with the CALPUFF - buoyant area source
As expected, for both source types, an insufficient number of parti model lead to a lower buoyancy flux compared to the point source
cles leads to discrepancies in the results, whereas the percent variation because of the heat losses. This in turn means that the plume rise phe
produced by the number of particles tends to decrease when the input nomenon is reduced, explaining why a variation in the source diameter
parameter becomes closer to the maximum number of particles has less influence on the pollutant dispersion.
considered. In particular, for scenario PARTIC8, involving the emission The obtained sensitivity indexes make it possible to identify whether
of 24′ 896′ 160 particles, the percent variability is approximately 1% in there is a linear relationship between the input and the output datum. As
all the selected receptors. Considering scenario PARTIC4, which implies reported by Yegnan et al. (2002), if the sensitivity coefficient is constant
the release of 2′ 989′ 440 particles, a maximum variability of 3% has been over a range of input parameters (i.e. if there is not a change in the
found. Considering the purpose of the first part of this paper, which aims sensitivity with a change in the input variable), the input datum can be
to investigate the model response to “macroscopic variations” in the considered to be linear with respect to the output.
input data, a “distorted” result of 3% is considered acceptable. Concerning CALPUFF, as shown in Fig. 5, the 3 lines are so close
together that they almost overlap, indicating an almost linear depen
dence of the ground level concentrations on the source diameter. Also,
3.4. Sensitivity analysis the distance between the curves increases moving away from the source,
and this means that this linear dependence is gradually reduced.
The development of alternative scenarios enables to identify the On the other hand, in SPRAY simulations (run with 106′ 688′ 880
most influential variables. Looking at the relative variations (%) ana particles), the lines seem to be more distant from each other. However,
lysed for each scenario in the previous paragraphs, the most critical the scattered behaviour does not allow to clearly identify the distances
input datum turns out to be the source diameter (except for the SPRAY - between the lines and, therefore, any comment concerning the input-
fire model). Although the amount of fuel burnt also significantly affects output relationship would be difficult.
the ground level concentrations, in the case of real fires, this is usually a Thus, to properly identify the relationship between the source
parameter that can be quantified with a certain degree of reliability. diameter and the pollutant concentration predicted by the two models,
That is why the sensitivity analysis, described in Section 2.8, has been some receptors, located at different distances from the source, have been
applied only to the source diameter. considered.
More in detail, the analysis has been applied to the point source Considering, for instance, receptors placed at a distance of 1000 m
simulated by SPRAY and the buoyant area source of CALPUFF, where it and 9000 m, respectively, from the source, the concentration trends
produces the main effects (see Section 3.2). predicted by SPRAY and CALPUFF when changing the source diameter
As already mentioned in Section 2.8.2, the source diameter has been are shown in Fig. 6:
perturbed by 1%, 2% and 3% in both directions with respect to the In addition, in each plot, the trend line (linear type) is drawn, and the
reference value of the base-case, while all the other parameters were kept resulting linear expression and the correlation coefficient (R2) are
unchanged. displayed.
Thus, 12 additional simulations, 6 for CALPUFF and 6 for SPRAY, Considering the receptor located at 1000 m from the source, it can be
have been run, setting the source diameter as reported in Table 6. stated that the linear model properly approximates the input-output
These additional runs allowed for comparison of the sensitivity of the relationship. This is true for both models, even if a higher R2 value is
SPRAY model (with 106′ 688′ 880 particles) and the CALPUFF model found for CALPUFF.
related to variations of the source diameter. By increasing the receptor distance from the source up to 9000 m, the
Looking at the trend of the normalized sensitivity index on the plume R2 coefficient decreases progressively for both models, even though a
axis (Fig. 5), a first comment concerns the sign of the coefficient. For better fit is still observed for the CALPUFF model. These considerations
CALPUFF, regardless of the considered receptor, the sensitivity coeffi are also confirmed by the analysis of other receptors, located at different
cient obtained by the puff model is always negative, revealing that a intermediate distances from the source (see Supplementary Material S5
negative correlation between the input and the output exists. The for receptors located at 3000 m, 5000 m and 7000 m from the source).
negative sign of the coefficient provides a clear indication that the The analysis on the individual receptors shows a decreasing reli
concentration value will be reduced because of an increment of the ability of the linear relationship between output and input values when
diameter. The same general behaviour is detected in SPRAY, except for moving away from the source. This may be justified considering the
few points, where a slightly positive value is obtained. wind effect on the dispersion phenomenon. Indeed, the Tracer concen
Another consideration concerns the general trend of the sensitivity trations modelled far from the source refer to the dispersion of the
coefficient as a function of the source distance. Both models produce the pollutant that has been subjected to the wind field for longer distances.
highest values (considering the absolute value) close to the source, The turbulent stochastic behaviour associated to the wind field pro
whereas values close to zero are identified for higher distances from the motes the plume distortion, which will be more pronounced far from the
source. This means that the investigated parameter has a greater influ emission source. This observation is also confirmed, for instance, by the
ence on the receptors close to the emission source. This can be ascribed ground level concentration maps resulting from the CALPUFF simula
to the plume rise mechanism, which is largely affected by the source tions when changing the source diameter from 5 m (base-case) to 10 m.
diameter: as previously discussed (Section 3.2), a change in the buoy The maps, reported in the Supplementary Material S6, show that, in the
ancy flux affects the pollutant dispersion close to the emission source vicinity of the source, the plume direction remains practically un
more than from far distances. changed, whereas at large distances, the wind effect results in a major
In addition, considering the receptors close to the source, higher plume deviation.
sensitivity coefficients are obtained when considering the point source Comparing the two models, it can be observed that, regardless of the
considered receptor, the R2 coefficients associated to the results of the
Table 6 SPRAY model are always lower than those of the CALPUFF model. The
Source diameter resulting from a variation of 1%, 2%, 3% of the average value. stochastic behaviour of the Lagrangian model influences the trajectory
Average value (base-case) = 5 m of each particle: the component related to the turbulent fluctuation
− 1% +1% − 2% +2% − 3% +3%
provides to each particle a random character. This chaotic contribution
leads to a loss of linearity between the input and the output variables.
4.95 m 5.05 m 4.90 m 5.10 m 4.85 m 5.15 m
To conclude, among all the information provided by this analysis, the
11
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Fig. 5. Normalized sensitivity index as a function of the source distance resulting from SPRAY (106′ 688′ 880 particles) and CALPUFF simulations by changing the
source diameter by 1%, 2% and 3%.
Fig. 6. Maximum 1-h Tracer concentration as a function of the source diameter predicted by the CALPUFF model (left) and by the SPRAY model (right) on two
receptors located at 1000 m and 9000 m from the source, respectively.
most interesting is the one related to the different sensitivities of SPRAY simulated as a point source and as a fire characterized by a 10% of
and CALPUFF to the investigated parameters. Indeed, the particle model emitted particles having no buoyancy flux.
appears to have a higher numerical sensitivity with respect to the The most relevant outcome resulting from the investigation of the
diameter of the emission source. This discrepancy is particularly evident alternative scenarios is that CALPUFF and SPRAY sensitivities to
in the vicinity of the emission source, where the sensitivity coefficients “macroscopic variations” of the considered parameters are generally
resulting from SPRAY simulations are higher by one order of magnitude. comparable. The only significant difference is the model sensitivity to
On the other hand, at larger distances the sensitivities of the two models the source diameter because, when using the specific option of the
appear comparable. SPRAY model to simulate emission from fires, the influence of the
diameter turns out to be negligible. Instead, the buoyant area source
4. Conclusions modelled by CALPUFF and the point source simulated by SPRAY are
significantly affected by variations in the estimation of the source
When modelling the environmental effects of atmospheric pollution, diameter. Indeed, the simulations conducted at the boundaries of the
many sources of imprecision and uncertainty affect the results and uncertainty range for the source diameter lead to variations of about
should therefore be critically analysed. Depending on the model 50–60% in the maximum Tracer concentrations at sensitive receptors.
considered, there are numerous potential sources of variability, such as On the other hand, when using SPRAY in combination with the specific
the input data required by the model. fire option, the source diameter variation gives a maximum variability of
This paper aimed to compare the sensitivity of the SPRAY and the 10%. These discrepancies can be explained considering the different
CALPUFF models to input parameters when simulating the pollutant relationships used to describe the plume rise phenomenon in the
dispersion from a hypothetical accidental fire, in order to identify the different models (see Section 3.2). It should be highlighted that the
most influential variables. In particular, the study focused on the effects choice of the emission model that best approximates the real physical
of input data regarding both source-term characterization and model- behaviour is essentially left to the user.
specific parameters. All the other investigated variables concerning the source term
To this purpose, starting from a reference “base-case” scenario, other characteristics, such as source height and emission temperature, do not
alternative emission scenarios, characterized by a “macroscopic varia significantly affect the model outputs, generally giving a maximum
tion” of each variable, have been defined. For each alternative scenario, variation in pollutant concentrations simulated at the receptors of about
three different source types have been studied: with CALPUFF the fire is 10%. The model-specific parameters point out a non-controlling influ
simulated as a buoyant area source, whereas with SPRAY the source is ence on the model results, generally producing variations of about 10%
12
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
13
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249
Henderson, S.B., Burkholder, B., Jackson, P.L., Brauer, M., Ichoku, C., 2008. Use of Ravina, M., Panepinto, D., Zanetti, M.C., 2020. Development of the DIDEM Model:
MODIS products to simplify and evaluate a forest fire plume dispersion model for comparative evaluation of CALPUFF and SPRAY dispersion models. Int. J. Environ.
PM10 exposure assessment. Atmos. Environ. 42, 8524–8532. https://doi.org/ Impacts Manag. Mitig. Recover. 3, 1–18. https://doi.org/10.2495/ei-v3-n1-1-18.
10.1016/j.atmosenv.2008.05.008. Rodrigues, H.S., Monteiro, M.T.T., Torres, D.F.M., 2013. Sensitivity analysis in a dengue
Holmes, N.S., Morawska, L., 2006. A review of dispersion modelling and its application epidemiological model. Conf. Pap. Math. 1–7. https://doi.org/10.1155/2013/
to the dispersion of particles: an overview of different dispersion models available. 721406, 2013.
Atmos. Environ. 40, 5902–5928. https://doi.org/10.1016/j.atmosenv.2006.06.003. Russell, A., Dennis, R., 2000. NARSTO critical review of photochemical models and
Holnicki, P., Kałuszko, A., Trapp, W., 2016. An urban scale application and validation of modeling. Atmos. Environ. 34, 2283–2324. https://doi.org/10.1016/S1352-2310
the CALPUFF model. Atmos. Pollut. Res. 7, 393–402. https://doi.org/10.1016/j. (99)00468-9.
apr.2015.10.016. Rzeszutek, M., 2019. Parameterization and evaluation of the CALMET/CALPUFF model
Holnicki, P., Nahorski, Z., 2015. Emission data uncertainty in urban air quality system in near-field and complex terrain - terrain data, grid resolution and terrain
modeling—case study. Environ. Model. Assess. 20, 583–597. https://doi.org/ adjustment method. Sci. Total Environ. 689, 31–46. https://doi.org/10.1016/j.
10.1007/s10666-015-9445-7. scitotenv.2019.06.379.
Ingason, H., Li, Y.Z., 2015. Tunnel Fire Dynamics. https://doi.org/10.1007/978-1-4939- Santiago, J.L., Martín, F., 2008. SLP-2D: a new Lagrangian particle model to simulate
2199-7. pollutant dispersion in street canyons. Atmos. Environ. 42, 3927–3936. https://doi.
Invernizzi, M., Brancher, M., Sironi, S., Capelli, L., Piringer, M., Schauberger, G., 2020. org/10.1016/j.atmosenv.2007.05.038.
Odour impact assessment by considering short-term ambient concentrations: a multi- Scire, J.S., Strimaitis, D.G., Yamartino, R.J., 2000. A user’s guide for the CALPUFF
model and two-site comparison. Environ. Int. 144, 105990. https://doi.org/ dispersion model. Earth Tech. Inc 521.
10.1016/j.envint.2020.105990. Seibert, P., 2000. Uncertainties in atmospheric dispersion modelling and source
Invernizzi, M., Tagliaferri, F., Sironi, S., Tinarelli, G., Capelli, L., 2021. Simulating determination. In: Procedings Informal Work. Meteorol. Model. Support CTBT,
pollutant dispersion from accidental fires with a focus on source characterization. Vienna, pp. 1–4.
J. Heal. Pollut. 11 https://doi.org/10.5696/2156-9614-11.30.210612. Seland, Oø, Iversen, T., 1999. A scheme for black carbon and sulphate aerosols tested in a
Islam, M.A., 1999. Application of a Gaussian Plume model to determine the location of hemispheric scale, Eulerian dispersion model. Atmos. Environ. 33, 2853–2879.
an unknown emission source. Water. Air. Soil Pollut. 112, 241–245. https://doi.org/ https://doi.org/10.1016/S1352-2310(98)00389-6.
10.1023/A:1005047321015. Shie, R.H., Chan, C.C., 2013. Tracking hazardous air pollutants from a refinery fire by
Jang, C.B., Choi, S.W., Baek, J.B., 2015. CFD modeling and fire damage analysis of jet fire applying on-line and off-line air monitoring and back trajectory modeling. J. Hazard
on hydrogen pipeline in a pipe rack structure. Int. J. Hydrogen Energy 40, Mater. 261, 72–82. https://doi.org/10.1016/j.jhazmat.2013.07.017.
15760–15772. https://doi.org/10.1016/j.ijhydene.2015.09.070. Sinha, N., Ghose, M.K., Singh, G., Srivastava, S., Sinha, I.N., 2004. Classification of air
Jung, Y.R., Park, W.G., Park, O.H., 2003. Pollution dispersion analysis using the puff pollution dispersion models : a critical review. Proc. Natl. Semin. Environ. Eng. with
model with numerical flow field data. Mech. Res. Commun. 30, 277–286. https:// Spec. Emphas. Min. Environ. 2004–19.
doi.org/10.1016/S0093-6413(03)00024-7. Sonnemans, P.J.M., Körvers, P.M.W., Pasman, H.J., 2010. Accidents in “normal”
Kota, S.H., Ying, Q., Zhang, Y., 2013. Simulating near-road reactive dispersion of gaseous operation - can you see them coming? J. Loss Prev. Process. Ind. 23, 351–366.
air pollutants using a three-dimensional Eulerian model. Sci. Total Environ. 454 https://doi.org/10.1016/j.jlp.2010.01.001.
(455), 348–357. https://doi.org/10.1016/j.scitotenv.2013.03.039. Souto, M.J., Souto, J.A., Pérez-Muñuzuri, V., Casares, J.J., Bermúdez, J.L., 2001.
Langmann, B., Duncan, B., Textor, C., Trentmann, J., van der Werf, G.R., 2009. A comparison of operational Lagrangian particle and adaptive puff models for plume
Vegetation fire emissions and their impact on air pollution and climate. Atmos. dispersion forecasting. Atmos. Environ. 35, 2349–2360. https://doi.org/10.1016/
Environ. 43, 107–116. https://doi.org/10.1016/j.atmosenv.2008.09.047. S1352-2310(00)00537-9.
Leelőssy, Á., Molnár, F., Izsák, F., Havasi, Á., Lagzi, I., Mészáros, R., 2014. Dispersion Srinivas, C.V., Hari Prasad, K.B.R.R., Naidu, C.V., Baskaran, R., Venkatraman, B., 2016.
modeling of air pollutants in the atmosphere: a review. Cent. Eur. J. Geosci. 6, Sensitivity analysis of atmospheric dispersion simulations by FLEXPART to the WRF-
257–278. https://doi.org/10.2478/s13533-012-0188-6. simulated meteorological predictions in a coastal environment. Pure Appl. Geophys.
Lemieux, P.M., Lutes, C.C., Santoianni, D.A., 2004. Emissions of organic air toxics from 173, 675–700. https://doi.org/10.1007/s00024-015-1104-z.
open burning: a comprehensive review. Progress in Energy and Combustion Science. Tinarelli, G., 2017. SPRAY 3 . 1 - General Description and User ’ S Guide.
https://doi.org/10.1016/j.pecs.2003.08.001. UNI 10796, 2009. Valutazione della dispersione in atmosfera di effluenti aeriformi. http:
Liu, T., Marlier, M.E., DeFries, R.S., Westervelt, D.M., Xia, K.R., Fiore, A.M., Mickley, L. //store.uni.com/catalogo/uni-10796-2000.
J., Cusworth, D.H., Milly, G., 2018. Seasonal impact of regional outdoor biomass UNI 10964, 2009. Guida alla selezione dei modelli matematici per la previsione di
burning on air pollution in three Indian cities: Delhi, Bengaluru, and Pune. Atmos. impatto sulla qualità dell ’ aria evaluation of air quality impact. http://store.uni.co
Environ. 172, 83–92. https://doi.org/10.1016/j.atmosenv.2017.10.024. m/catalogo/uni-10964-2001?josso_back_to=http://store.uni.com/josso-security-ch
Liu, X., Godbole, A., Lu, C., Michal, G., Venton, P., 2015. Optimisation of dispersion eck.php&josso_cmd=login_o.
parameters of Gaussian plume model for CO2 dispersion. Environ. Sci. Pollut. Res. US EPA, 1998. AP-42, “Compilation of Air Emission Factors” Chapter 1: External
22, 18288. https://doi.org/10.1007/s11356-015-5404-8, 18299. Combustion Sources. https://www.epa.gov/nscep.
Markatos, N.C., Christolis, C., Argyropoulos, C., 2009. Mathematical modeling of toxic US EPA, 1995. B-95, User’s Guide for the Industrial Source Complex (ISC3) Dispersion
pollutants dispersion from large tank fires and assessment of acute effects for fire Models - Volume Ii - Description of Model Algorithms, fifth ed. https://www.epa.
fighters. Int. J. Heat Mass Tran. 52, 4021–4030. https://doi.org/10.1016/j. gov/nscep.
ijheatmasstransfer.2009.03.039. Weichenthal, S., Van Rijswijk, D., Kulka, R., You, H., Van Ryswyk, K., Willey, J.,
Mishra, K.B., Wehrstedt, K.D., 2015. Underground gas pipeline explosion and fire: CFD Dugandzic, R., Sutcliffe, R., Moulton, J., Baike, M., White, L., Charland, J.P.,
based assessment of foreseeability. J. Nat. Gas Sci. Eng. 24, 526–542. https://doi. Jessiman, B., 2015. The impact of a landfill fire on ambient air quality in the north: a
org/10.1016/j.jngse.2015.04.010. case study in Iqaluit, Canada. Environ. Res. 142, 46–50. https://doi.org/10.1016/j.
Nivolianitou, Z., Konstandinidou, M., Michalis, C., 2006. Statistical analysis of major envres.2015.06.018.
accidents in petrochemical industry notified to the major accident reporting system Yegnan, A., Williamson, D.G., Graettinger, A.J., 2002. Uncertainty analysis in air
(MARS). J. Hazard Mater. 137, 1–7. https://doi.org/10.1016/j. dispersion modeling. Environ. Model. Software 17, 639–649. https://doi.org/
jhazmat.2004.12.042. 10.1016/S1364-8152(02)00026-9.
Novozhilov, V., 2001. Computational fluid dynamics modeling of compartment fires. Zheng, B., Chen, G., 2011. Storage tank fire accidents. Process Saf. Prog. 30, 291–293.
Prog. Energy Combust. Sci. 27, 611–666. https://doi.org/10.1016/S0360-1285(01)
00005-3.
14